Re: [PATCH V2 3/8] soc: qcom-geni-se: Add interconnect support to fix earlycon crash

Akash Asthana <akashast@xxxxxxxxxxxxxx> · Wed, 18 Mar 2020 16:27:11 +0530

Hi Evan

On 3/18/2020 12:38 AM, Evan Green wrote:
On Tue, Mar 17, 2020 at 3:58 AM Akash Asthana <akashast@xxxxxxxxxxxxxx> wrote:
Hi Matthias,

On 3/14/2020 2:14 AM, Matthias Kaehlcke wrote:
Hi Akash,

On Fri, Mar 13, 2020 at 06:42:09PM +0530, Akash Asthana wrote:
V1 patch@https://patchwork.kernel.org/patch/11386469/ caused SC7180 system
to reset at boot time.
The v1 patch isn't relevant in the commit message, please just describe the
problem. Also the crash only occurs when earlycon is used.
ok
As QUP core clock is shared among all the SE drivers present on particular
QUP wrapper, the reset seen is due to earlycon usage after QUP core clock
is put to 0 from other SE drivers before real console comes up.

As earlycon can't vote for it's QUP core need, to fix this add ICC
support to common/QUP wrapper driver and put vote for QUP core from
probe on behalf of earlycon and remove vote during sys suspend.
Only removing the vote on suspend isn't ideal, the system might never get
suspended. That said I don't have a really good alternative suggestion.

One thing you could possibly do is to launch a delayed work, check
console_device() every second or so and remove the vote when it returns
non-NULL. Not claiming this would be a great solution ...

The cleanest solution might be a notifier when the early console is
unregistered, it seems somewhat over-engineered though ... Then again
other (future) uart drivers with interconnect support might run into
the same problem.
We are hitting this problem because QUP core clocks are shared among all
the SE driver present in particular QUP wrapper, if other HW controllers
has similar architecture we will hit this issue.

How about if we expose an API from common driver(geni-se) for putting
QUP core BW vote to 0.

We call this from console probe just after uart_add_one_port call
(console resources are enabled as part of this call) to put core quota
to 0 on behalf of earlyconsole?
+Georgi

Hm, these boot proxy votes are annoying, since the whole house of
cards comes down if you replace these votes in the wrong order.

I believe consensus in the other patches was to consolidate most of
the interconnect support into the common SE code, right?

I think what Matthias suggested is to maintain ICC functions defined 
across I2C, SPI and UART as a library in common SE code.

Still every SE driver will interact with ICC framework individually 
rather than using common SE driver as a bridge.

  Would that
help you with these boot proxy votes? What I'm thinking is something
along the lines of:
  * SPI, I2C, UART all call into the new common geni_se_icc_on/off()
(or whatever it's called)
  * If geni_se_icc_off() sees that console UART hasn't voted yet, save
the votes but don't actually call icc_set(0) now.
  * Once uart votes for the first time, call icc_set() on all of SPI,
I2C, UART to get things back in sync.

IIUC, you are suggesting to enhancing ICC 
design@https://patchwork.kernel.org/patch/10774897/ [The very first ICC 
patch posted during sdm845 timeframe].

Where common SE driver aggregate real time BW requirement from all the 
SE driver and put net request to ICC framework.

We received comments on that version of ICC to move voting to individual 
SE driver from common driver. Hence we updated the design accordingly.

Thanks for reviewing

regards,

Akash

That's a sort of roll-your-own solution for GENI, but we do have this
problem elsewhere as well. A more general solution would be to have
the interconnect providers prop things up (ie ignore votes to lower
bandwidth) until some "go" moment where we feel we've enumerated all
devices. I was originally thinking to model this off of something like
clk_disable_unused(), but after chatting with Stephen it's clear
late_initcall's aren't really indicative of all devices having
actually come up. So I'm not sure where the appropriate "go" moment
is.

-Evan


Signed-off-by: Akash Asthana <akashast@xxxxxxxxxxxxxx>
Reported-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
---
   drivers/soc/qcom/qcom-geni-se.c | 41 +++++++++++++++++++++++++++++++++++++++++
   1 file changed, 41 insertions(+)

diff --git a/drivers/soc/qcom/qcom-geni-se.c b/drivers/soc/qcom/qcom-geni-se.c
index 7d622ea..d244dfc 100644
--- a/drivers/soc/qcom/qcom-geni-se.c
+++ b/drivers/soc/qcom/qcom-geni-se.c
@@ -90,6 +90,7 @@ struct geni_wrapper {
      struct device *dev;
      void __iomem *base;
      struct clk_bulk_data ahb_clks[NUM_AHB_CLKS];
+    struct icc_path *icc_path_geni_to_core;
   };

   #define QUP_HW_VER_REG                     0x4
@@ -747,11 +748,50 @@ static int geni_se_probe(struct platform_device *pdev)
              }
      }

+#ifdef CONFIG_SERIAL_EARLYCON
+    wrapper->icc_path_geni_to_core = devm_of_icc_get(dev, "qup-core");
+    if (IS_ERR(wrapper->icc_path_geni_to_core))
+            return PTR_ERR(wrapper->icc_path_geni_to_core);
+    /*
+     * Put minmal BW request on core clocks on behalf of early console.
+     * The vote will be removed in suspend call.
+     */
+    ret = icc_set_bw(wrapper->icc_path_geni_to_core, Bps_to_icc(1000),
+                    Bps_to_icc(1000));
+    if (ret) {
+            dev_err(&pdev->dev, "%s: ICC BW voting failed for core\n",
+                    __func__);
+            return ret;
+    }
What is ugly about this is that it's done for every QUP, not only the one
with the early console. Again, I don't have a good solution for it, maybe
it's a limitation we have to live with :(
There is one more limitation from QUP core side. Core clocks for both
the QUP wrapper runs at same speed.

core2x_1 = core2x_2 = max(core2x_1, core2x_2);

So with above limitation and if we are removing early con vote from Core
when real console comes up. It doesn't matter whether it's done for
every QUP or the only with early console.

+#endif
+
      dev_set_drvdata(dev, wrapper);
      dev_dbg(dev, "GENI SE Driver probed\n");
      return devm_of_platform_populate(dev);
   }

+static int __maybe_unused geni_se_sys_suspend(struct device *dev)
+{
+    struct geni_wrapper *wrapper = dev_get_drvdata(dev);
+    int ret;
+
+#ifdef CONFIG_SERIAL_EARLYCON
+    ret = icc_set_bw(wrapper->icc_path_geni_to_core, 0, 0);
I think you only want to do this on the first suspend.
Ok, I can add that logic using global variable.
Do we need to handle the case where no 'real' console is configured?
In this case the early console would be active forever and setting
the bandwidths to 0 might cause a similar crash than the one you are
trying to fix. Not sure if that's a real world use case, but wanted to
mention it. Maybe this is an argument of the notifier approach?
We can't support earlycon without real console.

As earlyconsole doesn't do any kind of resource enablement(SE clocks,
pinctrl, etc) it assumes that resources are already enabled from
previous stages.

So if real console doesn't come up no one will vote for that SE clock,
and it will be disabled from clk late_init call which will result into
un-clocked access.


+    if (ret) {
+            dev_err(dev, "%s: ICC BW remove failed for core\n",
+                    __func__);
+            return ret;
Aborting suspend seems too harsh since the QUP should still be fully
functional unless there is a general problem with the interconnects.

I would suggest to change the log to dev_warn() and return 0.
Ok

Thanks for reviewing the patch.

regards,

Akash

--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project

--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project






